{"id":"707924d9-7f60-40e1-90b6-995d4f1f7bef","arxiv_id":"2506.07485","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Penalizing the terminal constraint and sending the penalty to infinity produces an optimal control for extended mean field games with terminal constraint, characterized by a new conditional mean field FBSDE with a free backward part.","lead":"This paper solves a mean field game where every player must end with exactly zero holdings, using a penalty method and a limit argument. The result gives an optimal strategy for optimal liquidation and exhaustible-resource models, and proves solvability of a new kind of conditional forward-backward stochastic differential equation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3.21) requires C3 ≥ C1 but Lemma 3.4 states C3 < C1, so the terminal blow-up argument proving X∞_T = 0 is algebraically invalid as written.","rationale":"The reader's weakest_assumption highlights the nonnegativity and monotonicity machinery that drives the uniform bounds and the terminal blow-up in Lemma 3.7. Our stress test sharpens this to a concrete algebraic defect in that blow-up step: inequality (3.21) is false under the paper's own constants. The issue is directly load-bearing because Theorem 2.1 constructs an optimal control α∞ whose admissibility requires X∞_T = 0, and this terminal condition is proved only via (3.21). Notably, the flaw is not merely a missing sign control on X∞ − E[X∞|F^W0]; even if that difference vanished, the chain C1(X−ν)+C3ν ≥ C1X would still require C3 ≥ C1, contradicting Lemma 3.4. This makes the proof gap concrete and locatable. We also reviewed the other concerns: the terminal-condition inconsistency in (3.3) appears to be a typo fixable by setting φ_T = 0, and the reliance on the unpublished preprint [17] in Lemma 3.5 is a credibility concern rather than a stated internal contradiction. Because the penalization approach is plausible and the error is localized, a conditional verdict remains appropriate: the paper should not be rejected outright, but the authors must repair or replace (3.21) before the central claim is supported. Since the reader already assigned CONDITIONAL, our stress test does not change the verdict.","tokens_in":24302,"tokens_out":16415,"duration_ms":194268,"concrete_test":"Recompute (3.21) using only the bounds in Lemma 3.4. With P∞ ≥ C1/(T−t) and Ψ∞(t,ν) := P∞(t)ν + Φ∞(t,ν) ≥ C3 ν/(T−t), derive the claimed lower bound u∞(t,x,ν) ≥ C1 x/(T−t). The derivation reduces to requiring C1(x−ν) + C3ν ≥ C1x, i.e., (C3−C1)ν ≥ 0. Since X∞ ≥ 0 and ν = E[X∞|F^W0] can be strictly positive, this fails exactly when C3 < C1. Alternatively, evaluate a minimal one-dimensional instance (A=0, B=R=Q=1, f=b=l=h=0, ξ=1) where the Riccati solution and u∞ are explicit; check whether the claimed chain holds for the paper's constants. If the chain fails, the terminal blow-up step is unsupported and X∞_T = 0 is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorem 2.1) requires the limiting controlled state to satisfy X∞_T = 0. The only proof of this terminal condition is the blow-up argument in Lemma 3.7, which rests on inequality (3.21): u∞ ≥ C1(X∞ − E[X∞|F^W0])/(T−t) + C3 E[X∞|F^W0]/(T−t) ≥ C1 X∞/(T−t). The first lower bound is obtained by passing to the limit in the estimates P^L ≥ C1/(1/L+T−t) and P^L ν + Φ^L(ν) ≥ C3 ν/(1/L+T−t). But Lemma 3.4 explicitly states that C3 < C1. Since E[X∞|F^W0] ≥ 0 by nonnegativity of X∞, the second inequality in (3.21) is equivalent to (C3 − C1) E[X∞|F^W0]/(T−t) ≥ 0, which is false whenever the conditional mean is positive. Thus (3.21) is invalid as written, and the subsequent inequality X∞_t ≤ (tK+1)ξ+tKE[ξ] − (C1δ²/K)∫_0^t X∞_s/(T−s)ds does not follow. Without X∞_T = 0, the control α∞ defined by (3.22) is not admissible for problem (C-MF), and the equality V∞(ξ) = V(ξ) in Theorem 2.1 lacks its key step. This is not a minor typographical issue: the ordering C3 < C1 is essential in Lemma 3.4 because Φ_ν can be negative, so the constants cannot simply be swapped.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies a linear-quadratic extended mean field game with common noise and a strict terminal state constraint X_T = 0. The proposed method penalizes the terminal constraint by adding L X_T^2 to the cost, solves the penalized problem through a stochastic maximum principle and a conditional mean field FBSDE, and then analyzes the limiting behavior of the decoupling field u^L and the state X^L as L tends to infinity. The main result, Theorem 2.1, asserts that the limit control α^∞ is optimal for the constrained problem, and Theorems 3.8–3.9 characterize the limit as the unique solution of a new type of conditional mean field FBSDE in which the terminal condition is imposed on the forward component while the backward component has no boundary condition.","tokens_in":24677,"tokens_out":15000,"duration_ms":153773,"significance":"The paper addresses a genuinely difficult extension of the decoupling-field method: the common noise makes the decoupling field a function of both the spatial and the measure variable, and the terminal constraint produces a singular limiting object. If the proof can be completed, the main contribution is a constructive solution of a terminal-constrained extended MFG and a new wellposedness result for free-boundary conditional mean field FBSDEs. I also credit the authors for not assuming the main theorem: the limit objects u^∞ and X^∞ are defined as limits of explicitly constructed penalized objects, and Theorem 3.9 provides a uniqueness statement. However, the submitted manuscript contains a terminal-condition inconsistency in Section 3.1 and an invalid inequality in the key blow-up argument of Lemma 3.7, so the central claims are not presently established.","major_comments":[{"comment":"The terminal condition φ^L_T = L ν^L_T is inconsistent with the decoupling-field representation. For FBSDE (3.1) one has Y^L_T = L X^L_T; if Y^L_t = P^L_t X^L_t + Φ^L(t, E[X^L_t | F^{W^0}_t]) and P^L_T = L, then necessarily Φ^L(T,ν)=0 for every ν. In contrast, the definition Φ^L(t,ν)=φ^{t,ν}_t together with (3.3) gives Φ^L(T,ν)=Lν, so u^L(T,x,ν)=L(x+ν), not Lx. This mismatch concerns the wellposedness statement in Theorem 3.1 and the identification of the decoupling field, and it should be fixed (for example, by using φ^L_T = 0 in (3.3) or by redefining Φ^L and (3.7)).","section":"3.1, Eq. (3.3) and (3.4)"},{"comment":"The inequality chain in Eq. (3.21) is algebraically invalid. From Lemma 3.4, after L→∞ the lower bounds give u^∞(s,x,ν) ≥ C1 (x−ν)/(T−s) + C3 ν/(T−s). Since Lemma 3.4 states C3 < C1, the right-hand side equals C1 x/(T−s) + (C3−C1)ν/(T−s), which is strictly smaller than C1 x/(T−s) whenever ν = E[X^∞_s | F^{W^0}_s] > 0. Thus the claimed second inequality in (3.21) fails, and the subsequent estimate leading to X^∞_T = 0 is not established. The terminal constraint in Theorem 2.1 and the equality V^∞(ξ)=V(ξ) rest on this step; the ordering C3 < C1 is not an arbitrary typo because Φ^L_ν can be negative.","section":"3.2, Lemma 3.7, Eq. (3.21)"},{"comment":"The derivations of the dynamics of Ψ^L and Λ^L are delegated to 'similar arguments as [17]', a preprint by the same authors, and Theorem 3.1 is imported from [18]. These results are load-bearing: they supply the lower bounds (3.5)–(3.6) and the monotonicity in L that are used to define u^∞. The manuscript should either reproduce the arguments or state the precise results from [17] with matching hypotheses; as submitted, the proof is not self-contained at a critical point.","section":"3.1–3.2, Lemmas 3.4 and 3.5"},{"comment":"The comparison argument for γ_t = ∂u/∂L asserts that P_t and Λ_t are uniformly bounded 'according to Lemma 3.4'; however, Lemma 3.4 only provides bounds of order 1/(T−t) (κ_t and κ̃_t), so uniformity on [0,T) is not available. The proof needs to explain why the linear BSDE for γ_t has bounded coefficients on each interval used for the extension, or otherwise justify the comparison and the monotonicity of u^L in L.","section":"3.2, Lemma 3.5"}],"minor_comments":[{"comment":"The displayed definition 'K_3 = K^2/δ + (K^2+K^3)/(δ ε_0)' is circular; the constants should be reindexed to avoid using the same symbol on both sides.","section":"3.2, proof of Lemma 3.4"},{"comment":"The sentence that 'u^L is defined on a closed and bounded interval' is inaccurate: Lemma 3.2 concerns the state X^L, not u^L, and the domain of u^L is [0,T)×(0,∞)×(0,∞). The Arzelà–Ascoli step should be stated on compact subintervals of [0,T) and on bounded subsets of (0,∞)^2.","section":"3.2, proof of Lemma 3.6"},{"comment":"The phrase 'pointwise convergence of u^L(t,x,L)' should read 'u^L(t,x,ν)'; using L for both the penalty parameter and an argument of u^L is confusing.","section":"3.2, after Eq. (3.16)"},{"comment":"The terminal condition of FBSDE (3.8) is written as (0∨X_T∧C)(0∨L_T∧C); the text should explicitly justify that under Lemma 3.2 and L>0 the relevant values are in [0,C], so that this indeed coincides with X_T L_T.","section":"3.2, statement of Lemma 3.5"},{"comment":"The theorem is formulated for 0≤t≤r<T, but the displayed FBSDE (3.25) is written as a system over the whole interval with X^∞_T = 0; please clarify the sense in which (3.25) holds and whether Y^∞ is required to have a terminal value.","section":"3.3, Theorem 3.8"}],"recommendation":"major_revision","confidential_remarks":"The main theorem is plausible but the paper is not ready in its present form. The most serious issue is the invalid inequality (3.21); if the authors can supply a correct argument for X^∞_T = 0, the paper could be suitable. The editor may also want to verify that the results claimed from [17] are available and correct, since that preprint is not peer-reviewed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThe headline: this paper takes a promising route to terminal-constrained extended MFGs—penalize the terminal state, get a decoupling field for the penalized conditional mean field FBSDE, then pass L to infinity—but the proof as submitted does not close. The limiting blow-up argument in Lemma 3.7 contains an algebraic falsehood, and the terminal condition in the auxiliary FBSDE (3.3) is inconsistent with the decoupling field ansatz.\n\nWhat is genuinely new: the paper is the first to use decoupling fields for extended MFGs with a strict terminal state constraint and common noise. It removes the weak interaction condition from Fu, Graewe, Horst, and Popier by imposing a monotonicity/absorption structure, and it identifies the limiting object as a new type of conditional mean field FBSDE with two constraints on the forward component. The organization is clear, and the high-level strategy is coherent.\n\nThe soft spots are substantive. Eq. (3.21) asserts\nu∞ ≥ C1(X∞−E[X∞])/(T−t) + C3 E[X∞]/(T−t) ≥ C1 X∞/(T−t).\nThe second inequality requires C3 ≥ C1. Lemma 3.4 explicitly states C3 < C1 and its proof depends on that ordering. Unless the conditional mean is zero a.s.—which the framework does not guarantee—the chain is false. This is not a typo; the subsequent differential inequality and the conclusion X∞_T = 0 depend on it. Without X∞_T = 0, the limiting control is not admissible for (C-MF), and the equality V∞ = V in Theorem 2.1 is unsupported.\n\nSecond, Eq. (3.3) has φ^L_T = Lν^L_T, which gives Φ^L(T,ν)=Lν for the decoupling field. But FBSDE (3.1) has Y_T = LX_T, and the ansatz u^L(T,x,ν)=P^L_T x + Φ^L(T,ν) forces Φ^L(T,ν)=0. The two terminal conditions are incompatible. This looks like a modeling error in the auxiliary FBSDE rather than a harmless convention.\n\nThird, Lemmas 3.4 and 3.5 lean on “similar arguments as [17]”, a preprint by the same authors. That may be fine in a working paper, but a referee cannot verify the dynamics of Ψ^L or the global extension from those references without more detail.\n\nAll three issues are locatable and plausibly repairable. The core idea is good, and the paper deserves a serious referee. But the current version is not publishable as is. I would send it to peer review with a strong request for major revision, asking the authors to fix the constants in Lemma 3.7, reconcile the terminal conditions, and expand the arguments imported from [17].\n\nFor your reading group: maybe—worth reading for the method, but warn people about the gaps.","headline":"Promising approach to terminal-constrained extended MFGs via penalized decoupling fields, but the submitted proof has an invalid inequality in the main blow-up argument and an inconsistent terminal condition, so the result is not yet established.","tokens_in":25205,"tokens_out":3970,"would_cite":false,"duration_ms":39029,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E20","60H30","49N70"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that a terminal-constrained extended mean field game can be solved by adding a quadratic penalty on the terminal state and letting the penalty weight go to infinity, passing through the monotonicity of the associated…","keywords":["extended mean field game","terminal constraint","common noise","conditional mean field FBSDE","decoupling field","penalization method","optimal liquidation","absorption"],"falsifier":"Compute the limit X^\\infty from (3.20) in the one-dimensional pure liquidation case (A=0, B=1, f=b=l=h=0, Q=R=1) with positive deterministic \\xi. If X^\\infty_T>0, the terminal constraint is missed and Theorem 2.1 fails; alternatively, any concrete (t,x,\\nu) with x,\\nu>0 where u^L(t,x,\\nu) is not monotone in L falsifies Lemma 3.5.","tokens_in":24082,"feed_emoji":"🎯","tokens_out":7453,"duration_ms":66518,"temperature":0.7,"pith_summary":"This paper treats a class of linear-quadratic extended mean field games with common noise in which every player's state must be exactly zero at the terminal time, a constraint that makes the value function singular at T. The authors' plan is to replace the constrained game by an unconstrained one with a quadratic penalty on the terminal state, and then to let the penalty weight L tend to infinity. The core step is proving that the decoupling field u^L(t,x,\\nu)=P^L_t x+\\Phi^L(t,\\nu) of the associated conditional mean field forward-backward SDE is uniformly bounded and monotone in L on positive states; this yields limiting objects u^\\infty and X^\\infty with X^\\infty_T=0. They then identify the limit as the unique solution of a new type of coupled conditional mean field FBSDE in which the forward component carries both initial and terminal conditions and the backward component has no boundary condition. A sympathetic reader would care because the result gives a probabilistic, FBSDE-based route to terminal-constrained mean field equilibria without a weak-interaction assumption.","feed_headline":"Penalization limit solves terminal-constrained mean field games","feed_subtitle":"Adding a growing quadratic penalty and letting it blow up yields an optimal strategy whose state lands exactly at zero.","key_machinery":"The workhorse is the decoupling field u^L(t,x,\\nu)=P^L_t x+\\Phi^L(t,\\nu) of the penalized conditional mean field FBSDE (3.1), where P^L_t solves a Riccati ODE with terminal value L and \\Phi^L is the unique Lipschitz solution of an auxiliary FBSDE driven by the common noise. The field encodes the optimal cost-to-go for the unconstrained problem and provides the feedback form of the optimal control. Its monotonicity in L (Lemma 3.5) and uniform upper and lower bounds on positive states (Lemma 3.4) allow the limit L\\to\\infty to be taken, producing u^\\infty and the limiting state X^\\infty; the lower bound u^\\infty(s,X^\\infty_s,\\cdot)\\ge C_1 X^\\infty_s/(T-s) then drives the terminal blow-up argument that forces X^\\infty_T=0.","core_discovery":"The central discovery is that the constrained extended MFG has an optimal control \\$\\alpha$^\\infty, given in feedback form by \\$\\alpha$^\\infty_t=-$B_tR_t^{{-1}}$u^\\infty(t,X^\\infty_t,E[X^\\infty_t|$F^{{W^0}}$_t])-h(t,\\rho(t,-$R_t^{{-1}}$B_tu^\\infty(t,X^\\infty_t,E[X^\\infty_t|$F^{{W^0}}$_t])))) for t\\in[0,T) and \\$\\alpha$^\\infty_T=0, and that the associated quadruplet (X^\\infty,Y^\\infty,Z^\\infty,$Z^{{0,\\infty}}$) is the unique solution of the conditional mean field FBSDE (3.25). The control is obtained as the L\\to\\infty limit of the penalized optimal controls; the limit exists because the decoupling field u^L is non-decreasing in L and uniformly bounded for positive states, and the penalized optimal paths are nonnegative and non-increasing, which ultimately forces X^\\infty_T=0 through the blow-up estimate (3.21). In the limiting FBSDE the forward process carries both X^\\infty_0=\\xi and X^\\infty_T=0, while the backward component Y^\\infty satisfies no boundary condition; this is described as a new type of coupled conditional mean field FBSDE whose wellposedness is established in Theorem 3.9.","pith_inferences":["This construction suggests a numerical recipe: solve the penalized FBSDE for a sequence of growing L and use the monotone convergence of u^L to extrapolate X^L_T toward zero; a natural test is the pure liquidation case with A=0, B=1, Q=R=1.","Because the limiting backward equation carries no terminal data, the same decoupling-field limit may apply to other problems with hard end-point targets, such as principal-agent contracts with terminal capital requirements, where the free backward component plays the role of a shadow price.","The proof of monotonicity of u^L in L is carried out on strictly positive states; whether the convergence extends to states that reach zero before T, i.e. along the absorption boundary, is a boundary question the paper does not settle and could be probed numerically."],"forward_implications":["The optimal strategy \\alpha^\\infty is implementable as a feedback control depending on the state and its conditional expectation with respect to common noise, with the terminal value set to zero, and it attains finite cost for the constrained problem.","The penalized problems are consistent with the constrained problem: V(\\xi)=\\lim_{L\\to\\infty}V^L(\\xi), so solving the limit of unconstrained games gives the true value of the constrained game.","The theorem establishes wellposedness of a coupled conditional mean field FBSDE where the forward component has both an initial and a terminal condition and the backward component has no boundary condition at all; this is a new object even outside the MFG context.","The result covers nonlinear dependence of the coefficients on the mean field of states and controls, extending earlier terminal-constrained MFG results that required linear structure or weak interaction."],"supporting_citations":[{"why":"Provides the wellposedness of the conditional mean field FBSDE (3.1), the Riccati equation, and the optimal feedback control for the unconstrained penalized problem (P-MF).","marker":"[18]"},{"why":"Supplies the decoupling-field method for control problems with terminal state constraints that the present paper extends to extended MFGs with common noise.","marker":"[3]"},{"why":"Gives the probabilistic theory of conditional mean field FBSDEs under common noise and the small-time uniqueness result used to propagate monotonicity of u^L.","marker":"[8]"},{"why":"Provides the global solvability criterion for FBSDEs with diagonal generators and the dynamics of the derivative field used in the uniform bounds of Lemma 3.4.","marker":"[17]"},{"why":"Sets out the prior terminal-constrained MFG of optimal portfolio liquidation whose linear or weak-interaction hypotheses the paper removes via the monotonicity condition.","marker":"[10]"},{"why":"Motivates the absorption structure (states that hit zero drop out), which is the reason the optimal penalized paths stay nonnegative and non-increasing.","marker":"[11]"}],"fun_headline_variants":["Terminal-constrained MFGs solved via penalty limit","Penalty blow-up yields exact terminal state in MFGs","New FBSDE for mean field games with terminal constraints","How to enforce terminal state in mean field games: penalty limit","Decoupling fields crack terminal-constrained mean field games"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the absorption and monotonicity structure that keeps every optimal penalized state nonnegative and non-increasing — nonnegative initial state, A_t\\le 0, f'\\le 0, f(t,0)=b(t,0)=0 and the sign condition (2.1) — because without it the penalized trajectories need not converge to a path hitting zero at the terminal time.","fun_headline_variants_meta":{"raw":{"variants":["Terminal-constrained MFGs solved via penalty limit","Penalty blow-up yields exact terminal state in MFGs","New FBSDE for mean field games with terminal constraints","How to enforce terminal state in mean field games: penalty limit","Decoupling fields crack terminal-constrained mean field games"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000469,"raw_usage":{"total_tokens":2341,"prompt_tokens":959,"completion_tokens":1382,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":1302}},"tokens_in":575,"tokens_out":1382,"duration_ms":10513,"temperature":1.0,"reasoning_tokens":1302,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:35:38.300521+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the limit X^\\infty from (3.20) in the one-dimensional pure liquidation case (A=0, B=1, f=b=l=h=0, Q=R=1) with positive deterministic \\xi. If X^\\infty_T>0, the terminal constraint is missed and Theorem 2.1 fails; alternatively, any concrete (t,x,\\nu) with x,\\nu>0 where u^L(t,x,\\nu) is not monotone in L falsifies Lemma 3.5.","supporting_citations":[{"cited_title":"Hua and P","cited_arxiv_id":null,"evidence_quote":"Provides the wellposedness of the conditional mean field FBSDE (3.1), the Riccati equation, and the optimal feedback control for the unconstrained penalized problem (P-MF)."},{"cited_title":"Ankirchner, A","cited_arxiv_id":null,"evidence_quote":"Supplies the decoupling-field method for control problems with terminal state constraints that the present paper extends to extended MFGs with common noise."},{"cited_title":"Carmona and F","cited_arxiv_id":null,"evidence_quote":"Gives the probabilistic theory of conditional mean field FBSDEs under common noise and the small-time uniqueness result used to propagate monotonicity of u^L."},{"cited_title":"A unified approach to global solvability for FBSDEs with diagonal generators","cited_arxiv_id":"2211.00913","evidence_quote":"Provides the global solvability criterion for FBSDEs with diagonal generators and the dynamics of the derivative field used in the uniform bounds of Lemma 3.4."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sets out the prior terminal-constrained MFG of optimal portfolio liquidation whose linear or weak-interaction hypotheses the paper removes via the monotonicity condition."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the absorption structure (states that hit zero drop out), which is the reason the optimal penalized paths stay nonnegative and non-increasing."}],"review_version":1}